I've initially thought of sending the spans from PostgreSQL since this is the usual behavior of tracing libraries.
However, this created a lot potential issues:
- Protocol support and differences between trace collectors. OpenTelemetry seems to use gRPC, others are using http and those will require additional libraries (plus gRPC support in C doesn't look good) and any change in protobuf definition would require updating the extension.
- Do we send the spans within the query hooks? This means that we could block the process if the trace collector is slow to answer or we can’t connect. Sending spans from a background process sounded rather complex and resource heavy.
Moving to a pull model fixed those issues and felt more natural as this is the way PostgreSQL exposes its metrics.
> This patch looks very interesting, I'm working on the same subject too. But I've used > another approach - I'm using C wrapper for C++ API library from OpenTelemetry, and > handle span storage and output to this library. There are some nuances though, but it > is possible. Have you tried to use OpenTelemetry APIs instead of implementing all > functionality around spans?
I don't think that PostgreSQL accepts such kind of C++ code, not to mention the fact that the PostgreSQL license is not necessarily compatible with Apache 2.0 (I'm not a lawyer; this is not a legal advice). Such a design decision will probably require using separate compile flags since the user doesn't necessarily have a corresponding dependency installed. Similarly to how we do with LLVM, OpenSSL, etc.
So -1 to the OpenTelemetry C++ library and +1 to the properly licensed C implementation without 3rd party dependencies from me. Especially considering the fact that the implementation seems to be rather simple.
